照片影像的除霧及天空偵測演算法

天空是相當明顯的戶外照片特徵。我們提出一個高效率的天空偵測演算。首先，我們先找出一條較為粗略的天際線後，然後利用這條天際線求出天空機率模型需要的參數。最後我們使用天空機率模型決定各個像素屬於天空像素的機率。我們使用一個支援平行處理技術的影像處理函式庫實作所提出的演算法，在一個普通的桌面電腦上，VGA大小的圖片需要的處理時間小於35毫秒。也就是說，所提出的演算法在480P解析度的影片串流中可以實時地偵測天空。另一個和照片影像常常出現的問題是霾霧，有不少研究是關於如何除霧方法的。除霧演算法對很多應用都有助益，比如物件辨識。然而，除霧演算法會造成顏色失常。這些失常會出現在色彩的亮度和飽和度上，原因是除霧演算法做了一些猜測和補償那些被霾霧遮蔽的資訊。除霧效果越強，則色彩失常越大。我們提出一個即使使用很強的除霧效果，也可以保留色彩的架構，目標是讓輸入圖和結果圖的色調盡量相似。實驗結果表明我們的方法可以有效消除除霧後圖片的塊狀雜訊和色彩過飽和問題。另外，我們也探討了一些可以用於決定霾霧影響程度的特徵。我們的系統使用SVM和這些特徵決定一張圖片受霾霧影響的嚴重程度。這樣的系統可以用於影像內容辨識或幫助其他的除霧演算法決定參數。

關鍵字

實時天空偵測；天空機率模型；除霧演算法；霾霧嚴重程度判定

並列摘要

Sky is one of the most significant subject matters commonly seen in outdoor photos. We propose a highly efficient sky detection algorithm. First we detect a rough sky-ground boundary. Second we calculate parameters related to the appearance of sky. Finally, we use these parameters on a probability model that indicates how possible a pixel is belong to sky. And an image processing library with parallel processing techniques is used to implement proposed algorithm. In a common desktop computer, a VGA size image only requires less than 35ms, which can be considered a threshold for real-time processing, e.g. proposed algorithm may process 480p video in real time. Another characteristic commonly presented on photo image is haze. There are a number of researches on images’ dehazing. Dehazing technique is especially useful on applications such as object recognition. However there is a tradeoff between strength of haze-removing and tones of color. If we want to remove haze as much as possible, we may sacrifice tones of color. This tradeoff usually occurs in intensity and saturation, caused by dehazing algorithms which may do some guesses about information hidden by haze. A framework has been proposed to handle the tradeoff between tones of color and strength of removing haze. We aim to make human feel the same color tones after processing. Experiment results show that our framework is efficient to remove blocky and over-saturation effects on dehazed images. Furthermore, some features are explored for a haze-degree classification system which employs SVM as the learning model. This system is suitable for image content recognition or helping adjust parameters needed by haze-removal algorithms.

並列關鍵字

real-time ； sky detection ； sky probability ； haze removal ； dehaze ； haze-degree classification system

參考文獻

[1] Shen, Y., & Wang, Q. (2013). Sky Region Detection in a Single Image for Autonomous Ground Robot Navigation. International Journal of Advanced Robotic Systems, 10, 362.

[3] Zafarifar, B., & Weda, H. (2008, January). Horizon detection based on sky-color and edge features. In Electronic Imaging 2008 (pp. 682220-682220). International Society for Optics and Photonics.

[5] Luo, J., & Etz, S. P. (2002). A physical model-based approach to detecting sky in photographic images. Image Processing, IEEE Transactions on, 11(3), 201-212.

[7] Sutter, M., Dürr, B., & Philipona, R. (2004). Comparison of two radiation algorithms for surface‐based cloud‐free sky detection. Journal of Geophysical Research: Atmospheres (1984–2012), 109(D17).

[9] He, K., Sun, J., & Tang, X. (2011). Single image haze removal using dark channel prior. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 33(12), 2341-2353.

國際替代計量

照片影像的除霧及天空偵測演算法

全文下載

主題瀏覽